fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  44                               struct fuse_page_desc *page_descs,
  45                               unsigned npages)
  46 {
  47         INIT_LIST_HEAD(&req->list);
  48         INIT_LIST_HEAD(&req->intr_entry);
  49         init_waitqueue_head(&req->waitq);
  50         refcount_set(&req->count, 1);
  51         req->pages = pages;
  52         req->page_descs = page_descs;
  53         req->max_pages = npages;
  54         __set_bit(FR_PENDING, &req->flags);
  55 }
  56
  57 static struct page **fuse_req_pages_alloc(unsigned int npages, gfp_t flags,
  58                                           struct fuse_page_desc **desc)
  59 {
  60         struct page **pages;
  61
  62         pages = kzalloc(npages * (sizeof(struct page *) +
  63                                   sizeof(struct fuse_page_desc)), flags);
  64         *desc = (void *) pages + npages * sizeof(struct page *);
  65
  66         return pages;
  67 }
  68
  69 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  70 {
  71         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  72         if (req) {
  73                 struct page **pages = NULL;
  74                 struct fuse_page_desc *page_descs = NULL;
  75
  76                 WARN_ON(npages > FUSE_MAX_MAX_PAGES);
  77                 if (npages > FUSE_REQ_INLINE_PAGES) {
  78                         pages = fuse_req_pages_alloc(npages, flags,
  79                                                      &page_descs);
  80                         if (!pages) {
  81                                 kmem_cache_free(fuse_req_cachep, req);
  82                                 return NULL;
  83                         }
  84                 } else if (npages) {
  85                         pages = req->inline_pages;
  86                         page_descs = req->inline_page_descs;
  87                 }
  88
  89                 fuse_request_init(req, pages, page_descs, npages);
  90         }
  91         return req;
  92 }
  93
  94 struct fuse_req *fuse_request_alloc(unsigned npages)
  95 {
  96         return __fuse_request_alloc(npages, GFP_KERNEL);
  97 }
  98 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  99
 100 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
 101 {
 102         return __fuse_request_alloc(npages, GFP_NOFS);
 103 }
 104
 105 static void fuse_req_pages_free(struct fuse_req *req)
 106 {
 107         if (req->pages != req->inline_pages)
 108                 kfree(req->pages);
 109 }
 110
 111 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
 112                             gfp_t flags)
 113 {
 114         struct page **pages;
 115         struct fuse_page_desc *page_descs;
 116         unsigned int npages = min_t(unsigned int,
 117                                     max_t(unsigned int, req->max_pages * 2,
 118                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
 119                                     fc->max_pages);
 120         WARN_ON(npages <= req->max_pages);
 121
 122         pages = fuse_req_pages_alloc(npages, flags, &page_descs);
 123         if (!pages)
 124                 return false;
 125
 126         memcpy(pages, req->pages, sizeof(struct page *) * req->max_pages);
 127         memcpy(page_descs, req->page_descs,
 128                sizeof(struct fuse_page_desc) * req->max_pages);
 129         fuse_req_pages_free(req);
 130         req->pages = pages;
 131         req->page_descs = page_descs;
 132         req->max_pages = npages;
 133
 134         return true;
 135 }
 136
 137 void fuse_request_free(struct fuse_req *req)
 138 {
 139         fuse_req_pages_free(req);
 140         kmem_cache_free(fuse_req_cachep, req);
 141 }
 142
 143 void __fuse_get_request(struct fuse_req *req)
 144 {
 145         refcount_inc(&req->count);
 146 }
 147
 148 /* Must be called with > 1 refcount */
 149 static void __fuse_put_request(struct fuse_req *req)
 150 {
 151         refcount_dec(&req->count);
 152 }
 153
 154 void fuse_set_initialized(struct fuse_conn *fc)
 155 {
 156         /* Make sure stores before this are seen on another CPU */
 157         smp_wmb();
 158         fc->initialized = 1;
 159 }
 160
 161 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 162 {
 163         return !fc->initialized || (for_background && fc->blocked);
 164 }
 165
 166 static void fuse_drop_waiting(struct fuse_conn *fc)
 167 {
 168         /*
 169          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 170          * provides a memory barrier mached with the one in fuse_wait_aborted()
 171          * to ensure no wake-up is missed.
 172          */
 173         if (atomic_dec_and_test(&fc->num_waiting) &&
 174             !READ_ONCE(fc->connected)) {
 175                 /* wake up aborters */
 176                 wake_up_all(&fc->blocked_waitq);
 177         }
 178 }
 179
 180 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 181                                        bool for_background)
 182 {
 183         struct fuse_req *req;
 184         int err;
 185         atomic_inc(&fc->num_waiting);
 186
 187         if (fuse_block_alloc(fc, for_background)) {
 188                 err = -EINTR;
 189                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 190                                 !fuse_block_alloc(fc, for_background)))
 191                         goto out;
 192         }
 193         /* Matches smp_wmb() in fuse_set_initialized() */
 194         smp_rmb();
 195
 196         err = -ENOTCONN;
 197         if (!fc->connected)
 198                 goto out;
 199
 200         err = -ECONNREFUSED;
 201         if (fc->conn_error)
 202                 goto out;
 203
 204         req = fuse_request_alloc(npages);
 205         err = -ENOMEM;
 206         if (!req) {
 207                 if (for_background)
 208                         wake_up(&fc->blocked_waitq);
 209                 goto out;
 210         }
 211
 212         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 213         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 214         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 215
 216         __set_bit(FR_WAITING, &req->flags);
 217         if (for_background)
 218                 __set_bit(FR_BACKGROUND, &req->flags);
 219
 220         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 221                      req->in.h.gid == ((gid_t)-1))) {
 222                 fuse_put_request(fc, req);
 223                 return ERR_PTR(-EOVERFLOW);
 224         }
 225         return req;
 226
 227  out:
 228         fuse_drop_waiting(fc);
 229         return ERR_PTR(err);
 230 }
 231
 232 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 233 {
 234         return __fuse_get_req(fc, npages, false);
 235 }
 236 EXPORT_SYMBOL_GPL(fuse_get_req);
 237
 238 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 239                                              unsigned npages)
 240 {
 241         return __fuse_get_req(fc, npages, true);
 242 }
 243 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 244
 245 /*
 246  * Return request in fuse_file->reserved_req.  However that may
 247  * currently be in use.  If that is the case, wait for it to become
 248  * available.
 249  */
 250 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 251                                          struct file *file)
 252 {
 253         struct fuse_req *req = NULL;
 254         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 255         struct fuse_file *ff = file->private_data;
 256
 257         do {
 258                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 259                 spin_lock(&fi->lock);
 260                 if (ff->reserved_req) {
 261                         req = ff->reserved_req;
 262                         ff->reserved_req = NULL;
 263                         req->stolen_file = get_file(file);
 264                 }
 265                 spin_unlock(&fi->lock);
 266         } while (!req);
 267
 268         return req;
 269 }
 270
 271 /*
 272  * Put stolen request back into fuse_file->reserved_req
 273  */
 274 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 275 {
 276         struct file *file = req->stolen_file;
 277         struct fuse_inode *fi = get_fuse_inode(file_inode(file));
 278         struct fuse_file *ff = file->private_data;
 279
 280         WARN_ON(req->max_pages);
 281         spin_lock(&fi->lock);
 282         memset(req, 0, sizeof(*req));
 283         fuse_request_init(req, NULL, NULL, 0);
 284         BUG_ON(ff->reserved_req);
 285         ff->reserved_req = req;
 286         wake_up_all(&fc->reserved_req_waitq);
 287         spin_unlock(&fi->lock);
 288         fput(file);
 289 }
 290
 291 /*
 292  * Gets a requests for a file operation, always succeeds
 293  *
 294  * This is used for sending the FLUSH request, which must get to
 295  * userspace, due to POSIX locks which may need to be unlocked.
 296  *
 297  * If allocation fails due to OOM, use the reserved request in
 298  * fuse_file.
 299  *
 300  * This is very unlikely to deadlock accidentally, since the
 301  * filesystem should not have it's own file open.  If deadlock is
 302  * intentional, it can still be broken by "aborting" the filesystem.
 303  */
 304 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 305                                              struct file *file)
 306 {
 307         struct fuse_req *req;
 308
 309         atomic_inc(&fc->num_waiting);
 310         wait_event(fc->blocked_waitq, fc->initialized);
 311         /* Matches smp_wmb() in fuse_set_initialized() */
 312         smp_rmb();
 313         req = fuse_request_alloc(0);
 314         if (!req)
 315                 req = get_reserved_req(fc, file);
 316
 317         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 318         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 319         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 320
 321         __set_bit(FR_WAITING, &req->flags);
 322         __clear_bit(FR_BACKGROUND, &req->flags);
 323         return req;
 324 }
 325
 326 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 327 {
 328         if (refcount_dec_and_test(&req->count)) {
 329                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 330                         /*
 331                          * We get here in the unlikely case that a background
 332                          * request was allocated but not sent
 333                          */
 334                         spin_lock(&fc->bg_lock);
 335                         if (!fc->blocked)
 336                                 wake_up(&fc->blocked_waitq);
 337                         spin_unlock(&fc->bg_lock);
 338                 }
 339
 340                 if (test_bit(FR_WAITING, &req->flags)) {
 341                         __clear_bit(FR_WAITING, &req->flags);
 342                         fuse_drop_waiting(fc);
 343                 }
 344
 345                 if (req->stolen_file)
 346                         put_reserved_req(fc, req);
 347                 else
 348                         fuse_request_free(req);
 349         }
 350 }
 351 EXPORT_SYMBOL_GPL(fuse_put_request);
 352
 353 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 354 {
 355         unsigned nbytes = 0;
 356         unsigned i;
 357
 358         for (i = 0; i < numargs; i++)
 359                 nbytes += args[i].size;
 360
 361         return nbytes;
 362 }
 363
 364 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 365 {
 366         fiq->reqctr += FUSE_REQ_ID_STEP;
 367         return fiq->reqctr;
 368 }
 369
 370 static unsigned int fuse_req_hash(u64 unique)
 371 {
 372         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 373 }
 374
 375 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 376 {
 377         req->in.h.len = sizeof(struct fuse_in_header) +
 378                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 379         list_add_tail(&req->list, &fiq->pending);
 380         wake_up_locked(&fiq->waitq);
 381         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 382 }
 383
 384 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 385                        u64 nodeid, u64 nlookup)
 386 {
 387         struct fuse_iqueue *fiq = &fc->iq;
 388
 389         forget->forget_one.nodeid = nodeid;
 390         forget->forget_one.nlookup = nlookup;
 391
 392         spin_lock(&fiq->waitq.lock);
 393         if (fiq->connected) {
 394                 fiq->forget_list_tail->next = forget;
 395                 fiq->forget_list_tail = forget;
 396                 wake_up_locked(&fiq->waitq);
 397                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 398         } else {
 399                 kfree(forget);
 400         }
 401         spin_unlock(&fiq->waitq.lock);
 402 }
 403
 404 static void flush_bg_queue(struct fuse_conn *fc)
 405 {
 406         struct fuse_iqueue *fiq = &fc->iq;
 407
 408         while (fc->active_background < fc->max_background &&
 409                !list_empty(&fc->bg_queue)) {
 410                 struct fuse_req *req;
 411
 412                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 413                 list_del(&req->list);
 414                 fc->active_background++;
 415                 spin_lock(&fiq->waitq.lock);
 416                 req->in.h.unique = fuse_get_unique(fiq);
 417                 queue_request(fiq, req);
 418                 spin_unlock(&fiq->waitq.lock);
 419         }
 420 }
 421
 422 /*
 423  * This function is called when a request is finished.  Either a reply
 424  * has arrived or it was aborted (and not yet sent) or some error
 425  * occurred during communication with userspace, or the device file
 426  * was closed.  The requester thread is woken up (if still waiting),
 427  * the 'end' callback is called if given, else the reference to the
 428  * request is released
 429  */
 430 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 431 {
 432         struct fuse_iqueue *fiq = &fc->iq;
 433
 434         if (test_and_set_bit(FR_FINISHED, &req->flags))
 435                 goto put_request;
 436         /*
 437          * test_and_set_bit() implies smp_mb() between bit
 438          * changing and below intr_entry check. Pairs with
 439          * smp_mb() from queue_interrupt().
 440          */
 441         if (!list_empty(&req->intr_entry)) {
 442                 spin_lock(&fiq->waitq.lock);
 443                 list_del_init(&req->intr_entry);
 444                 spin_unlock(&fiq->waitq.lock);
 445         }
 446         WARN_ON(test_bit(FR_PENDING, &req->flags));
 447         WARN_ON(test_bit(FR_SENT, &req->flags));
 448         if (test_bit(FR_BACKGROUND, &req->flags)) {
 449                 spin_lock(&fc->bg_lock);
 450                 clear_bit(FR_BACKGROUND, &req->flags);
 451                 if (fc->num_background == fc->max_background) {
 452                         fc->blocked = 0;
 453                         wake_up(&fc->blocked_waitq);
 454                 } else if (!fc->blocked) {
 455                         /*
 456                          * Wake up next waiter, if any.  It's okay to use
 457                          * waitqueue_active(), as we've already synced up
 458                          * fc->blocked with waiters with the wake_up() call
 459                          * above.
 460                          */
 461                         if (waitqueue_active(&fc->blocked_waitq))
 462                                 wake_up(&fc->blocked_waitq);
 463                 }
 464
 465                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 466                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 467                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 468                 }
 469                 fc->num_background--;
 470                 fc->active_background--;
 471                 flush_bg_queue(fc);
 472                 spin_unlock(&fc->bg_lock);
 473         } else {
 474                 /* Wake up waiter sleeping in request_wait_answer() */
 475                 wake_up(&req->waitq);
 476         }
 477
 478         if (req->end)
 479                 req->end(fc, req);
 480 put_request:
 481         fuse_put_request(fc, req);
 482 }
 483
 484 static int queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 485 {
 486         spin_lock(&fiq->waitq.lock);
 487         /* Check for we've sent request to interrupt this req */
 488         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 489                 spin_unlock(&fiq->waitq.lock);
 490                 return -EINVAL;
 491         }
 492
 493         if (list_empty(&req->intr_entry)) {
 494                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 495                 /*
 496                  * Pairs with smp_mb() implied by test_and_set_bit()
 497                  * from request_end().
 498                  */
 499                 smp_mb();
 500                 if (test_bit(FR_FINISHED, &req->flags)) {
 501                         list_del_init(&req->intr_entry);
 502                         spin_unlock(&fiq->waitq.lock);
 503                         return 0;
 504                 }
 505                 wake_up_locked(&fiq->waitq);
 506                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 507         }
 508         spin_unlock(&fiq->waitq.lock);
 509         return 0;
 510 }
 511
 512 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 513 {
 514         struct fuse_iqueue *fiq = &fc->iq;
 515         int err;
 516
 517         if (!fc->no_interrupt) {
 518                 /* Any signal may interrupt this */
 519                 err = wait_event_interruptible(req->waitq,
 520                                         test_bit(FR_FINISHED, &req->flags));
 521                 if (!err)
 522                         return;
 523
 524                 set_bit(FR_INTERRUPTED, &req->flags);
 525                 /* matches barrier in fuse_dev_do_read() */
 526                 smp_mb__after_atomic();
 527                 if (test_bit(FR_SENT, &req->flags))
 528                         queue_interrupt(fiq, req);
 529         }
 530
 531         if (!test_bit(FR_FORCE, &req->flags)) {
 532                 /* Only fatal signals may interrupt this */
 533                 err = wait_event_killable(req->waitq,
 534                                         test_bit(FR_FINISHED, &req->flags));
 535                 if (!err)
 536                         return;
 537
 538                 spin_lock(&fiq->waitq.lock);
 539                 /* Request is not yet in userspace, bail out */
 540                 if (test_bit(FR_PENDING, &req->flags)) {
 541                         list_del(&req->list);
 542                         spin_unlock(&fiq->waitq.lock);
 543                         __fuse_put_request(req);
 544                         req->out.h.error = -EINTR;
 545                         return;
 546                 }
 547                 spin_unlock(&fiq->waitq.lock);
 548         }
 549
 550         /*
 551          * Either request is already in userspace, or it was forced.
 552          * Wait it out.
 553          */
 554         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 555 }
 556
 557 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 558 {
 559         struct fuse_iqueue *fiq = &fc->iq;
 560
 561         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 562         spin_lock(&fiq->waitq.lock);
 563         if (!fiq->connected) {
 564                 spin_unlock(&fiq->waitq.lock);
 565                 req->out.h.error = -ENOTCONN;
 566         } else {
 567                 req->in.h.unique = fuse_get_unique(fiq);
 568                 queue_request(fiq, req);
 569                 /* acquire extra reference, since request is still needed
 570                    after request_end() */
 571                 __fuse_get_request(req);
 572                 spin_unlock(&fiq->waitq.lock);
 573
 574                 request_wait_answer(fc, req);
 575                 /* Pairs with smp_wmb() in request_end() */
 576                 smp_rmb();
 577         }
 578 }
 579
 580 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 581 {
 582         __set_bit(FR_ISREPLY, &req->flags);
 583         if (!test_bit(FR_WAITING, &req->flags)) {
 584                 __set_bit(FR_WAITING, &req->flags);
 585                 atomic_inc(&fc->num_waiting);
 586         }
 587         __fuse_request_send(fc, req);
 588 }
 589 EXPORT_SYMBOL_GPL(fuse_request_send);
 590
 591 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 592 {
 593         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 594                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 595
 596         if (fc->minor < 9) {
 597                 switch (args->in.h.opcode) {
 598                 case FUSE_LOOKUP:
 599                 case FUSE_CREATE:
 600                 case FUSE_MKNOD:
 601                 case FUSE_MKDIR:
 602                 case FUSE_SYMLINK:
 603                 case FUSE_LINK:
 604                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 605                         break;
 606                 case FUSE_GETATTR:
 607                 case FUSE_SETATTR:
 608                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 609                         break;
 610                 }
 611         }
 612         if (fc->minor < 12) {
 613                 switch (args->in.h.opcode) {
 614                 case FUSE_CREATE:
 615                         args->in.args[0].size = sizeof(struct fuse_open_in);
 616                         break;
 617                 case FUSE_MKNOD:
 618                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 619                         break;
 620                 }
 621         }
 622 }
 623
 624 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 625 {
 626         struct fuse_req *req;
 627         ssize_t ret;
 628
 629         req = fuse_get_req(fc, 0);
 630         if (IS_ERR(req))
 631                 return PTR_ERR(req);
 632
 633         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 634         fuse_adjust_compat(fc, args);
 635
 636         req->in.h.opcode = args->in.h.opcode;
 637         req->in.h.nodeid = args->in.h.nodeid;
 638         req->in.numargs = args->in.numargs;
 639         memcpy(req->in.args, args->in.args,
 640                args->in.numargs * sizeof(struct fuse_in_arg));
 641         req->out.argvar = args->out.argvar;
 642         req->out.numargs = args->out.numargs;
 643         memcpy(req->out.args, args->out.args,
 644                args->out.numargs * sizeof(struct fuse_arg));
 645         fuse_request_send(fc, req);
 646         ret = req->out.h.error;
 647         if (!ret && args->out.argvar) {
 648                 BUG_ON(args->out.numargs != 1);
 649                 ret = req->out.args[0].size;
 650         }
 651         fuse_put_request(fc, req);
 652
 653         return ret;
 654 }
 655
 656 bool fuse_request_queue_background(struct fuse_conn *fc, struct fuse_req *req)
 657 {
 658         bool queued = false;
 659
 660         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 661         if (!test_bit(FR_WAITING, &req->flags)) {
 662                 __set_bit(FR_WAITING, &req->flags);
 663                 atomic_inc(&fc->num_waiting);
 664         }
 665         __set_bit(FR_ISREPLY, &req->flags);
 666         spin_lock(&fc->bg_lock);
 667         if (likely(fc->connected)) {
 668                 fc->num_background++;
 669                 if (fc->num_background == fc->max_background)
 670                         fc->blocked = 1;
 671                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 672                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 673                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 674                 }
 675                 list_add_tail(&req->list, &fc->bg_queue);
 676                 flush_bg_queue(fc);
 677                 queued = true;
 678         }
 679         spin_unlock(&fc->bg_lock);
 680
 681         return queued;
 682 }
 683
 684 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 685 {
 686         WARN_ON(!req->end);
 687         if (!fuse_request_queue_background(fc, req)) {
 688                 req->out.h.error = -ENOTCONN;
 689                 req->end(fc, req);
 690                 fuse_put_request(fc, req);
 691         }
 692 }
 693 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 694
 695 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 696                                           struct fuse_req *req, u64 unique)
 697 {
 698         int err = -ENODEV;
 699         struct fuse_iqueue *fiq = &fc->iq;
 700
 701         __clear_bit(FR_ISREPLY, &req->flags);
 702         req->in.h.unique = unique;
 703         spin_lock(&fiq->waitq.lock);
 704         if (fiq->connected) {
 705                 queue_request(fiq, req);
 706                 err = 0;
 707         }
 708         spin_unlock(&fiq->waitq.lock);
 709
 710         return err;
 711 }
 712
 713 void fuse_force_forget(struct file *file, u64 nodeid)
 714 {
 715         struct inode *inode = file_inode(file);
 716         struct fuse_conn *fc = get_fuse_conn(inode);
 717         struct fuse_req *req;
 718         struct fuse_forget_in inarg;
 719
 720         memset(&inarg, 0, sizeof(inarg));
 721         inarg.nlookup = 1;
 722         req = fuse_get_req_nofail_nopages(fc, file);
 723         req->in.h.opcode = FUSE_FORGET;
 724         req->in.h.nodeid = nodeid;
 725         req->in.numargs = 1;
 726         req->in.args[0].size = sizeof(inarg);
 727         req->in.args[0].value = &inarg;
 728         __clear_bit(FR_ISREPLY, &req->flags);
 729         __fuse_request_send(fc, req);
 730         /* ignore errors */
 731         fuse_put_request(fc, req);
 732 }
 733
 734 /*
 735  * Lock the request.  Up to the next unlock_request() there mustn't be
 736  * anything that could cause a page-fault.  If the request was already
 737  * aborted bail out.
 738  */
 739 static int lock_request(struct fuse_req *req)
 740 {
 741         int err = 0;
 742         if (req) {
 743                 spin_lock(&req->waitq.lock);
 744                 if (test_bit(FR_ABORTED, &req->flags))
 745                         err = -ENOENT;
 746                 else
 747                         set_bit(FR_LOCKED, &req->flags);
 748                 spin_unlock(&req->waitq.lock);
 749         }
 750         return err;
 751 }
 752
 753 /*
 754  * Unlock request.  If it was aborted while locked, caller is responsible
 755  * for unlocking and ending the request.
 756  */
 757 static int unlock_request(struct fuse_req *req)
 758 {
 759         int err = 0;
 760         if (req) {
 761                 spin_lock(&req->waitq.lock);
 762                 if (test_bit(FR_ABORTED, &req->flags))
 763                         err = -ENOENT;
 764                 else
 765                         clear_bit(FR_LOCKED, &req->flags);
 766                 spin_unlock(&req->waitq.lock);
 767         }
 768         return err;
 769 }
 770
 771 struct fuse_copy_state {
 772         int write;
 773         struct fuse_req *req;
 774         struct iov_iter *iter;
 775         struct pipe_buffer *pipebufs;
 776         struct pipe_buffer *currbuf;
 777         struct pipe_inode_info *pipe;
 778         unsigned long nr_segs;
 779         struct page *pg;
 780         unsigned len;
 781         unsigned offset;
 782         unsigned move_pages:1;
 783 };
 784
 785 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 786                            struct iov_iter *iter)
 787 {
 788         memset(cs, 0, sizeof(*cs));
 789         cs->write = write;
 790         cs->iter = iter;
 791 }
 792
 793 /* Unmap and put previous page of userspace buffer */
 794 static void fuse_copy_finish(struct fuse_copy_state *cs)
 795 {
 796         if (cs->currbuf) {
 797                 struct pipe_buffer *buf = cs->currbuf;
 798
 799                 if (cs->write)
 800                         buf->len = PAGE_SIZE - cs->len;
 801                 cs->currbuf = NULL;
 802         } else if (cs->pg) {
 803                 if (cs->write) {
 804                         flush_dcache_page(cs->pg);
 805                         set_page_dirty_lock(cs->pg);
 806                 }
 807                 put_page(cs->pg);
 808         }
 809         cs->pg = NULL;
 810 }
 811
 812 /*
 813  * Get another pagefull of userspace buffer, and map it to kernel
 814  * address space, and lock request
 815  */
 816 static int fuse_copy_fill(struct fuse_copy_state *cs)
 817 {
 818         struct page *page;
 819         int err;
 820
 821         err = unlock_request(cs->req);
 822         if (err)
 823                 return err;
 824
 825         fuse_copy_finish(cs);
 826         if (cs->pipebufs) {
 827                 struct pipe_buffer *buf = cs->pipebufs;
 828
 829                 if (!cs->write) {
 830                         err = pipe_buf_confirm(cs->pipe, buf);
 831                         if (err)
 832                                 return err;
 833
 834                         BUG_ON(!cs->nr_segs);
 835                         cs->currbuf = buf;
 836                         cs->pg = buf->page;
 837                         cs->offset = buf->offset;
 838                         cs->len = buf->len;
 839                         cs->pipebufs++;
 840                         cs->nr_segs--;
 841                 } else {
 842                         if (cs->nr_segs == cs->pipe->buffers)
 843                                 return -EIO;
 844
 845                         page = alloc_page(GFP_HIGHUSER);
 846                         if (!page)
 847                                 return -ENOMEM;
 848
 849                         buf->page = page;
 850                         buf->offset = 0;
 851                         buf->len = 0;
 852
 853                         cs->currbuf = buf;
 854                         cs->pg = page;
 855                         cs->offset = 0;
 856                         cs->len = PAGE_SIZE;
 857                         cs->pipebufs++;
 858                         cs->nr_segs++;
 859                 }
 860         } else {
 861                 size_t off;
 862                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 863                 if (err < 0)
 864                         return err;
 865                 BUG_ON(!err);
 866                 cs->len = err;
 867                 cs->offset = off;
 868                 cs->pg = page;
 869                 iov_iter_advance(cs->iter, err);
 870         }
 871
 872         return lock_request(cs->req);
 873 }
 874
 875 /* Do as much copy to/from userspace buffer as we can */
 876 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 877 {
 878         unsigned ncpy = min(*size, cs->len);
 879         if (val) {
 880                 void *pgaddr = kmap_atomic(cs->pg);
 881                 void *buf = pgaddr + cs->offset;
 882
 883                 if (cs->write)
 884                         memcpy(buf, *val, ncpy);
 885                 else
 886                         memcpy(*val, buf, ncpy);
 887
 888                 kunmap_atomic(pgaddr);
 889                 *val += ncpy;
 890         }
 891         *size -= ncpy;
 892         cs->len -= ncpy;
 893         cs->offset += ncpy;
 894         return ncpy;
 895 }
 896
 897 static int fuse_check_page(struct page *page)
 898 {
 899         if (page_mapcount(page) ||
 900             page->mapping != NULL ||
 901             page_count(page) != 1 ||
 902             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 903              ~(1 << PG_locked |
 904                1 << PG_referenced |
 905                1 << PG_uptodate |
 906                1 << PG_lru |
 907                1 << PG_active |
 908                1 << PG_reclaim))) {
 909                 pr_warn("trying to steal weird page\n");
 910                 pr_warn("  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 911                 return 1;
 912         }
 913         return 0;
 914 }
 915
 916 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 917 {
 918         int err;
 919         struct page *oldpage = *pagep;
 920         struct page *newpage;
 921         struct pipe_buffer *buf = cs->pipebufs;
 922
 923         err = unlock_request(cs->req);
 924         if (err)
 925                 return err;
 926
 927         fuse_copy_finish(cs);
 928
 929         err = pipe_buf_confirm(cs->pipe, buf);
 930         if (err)
 931                 return err;
 932
 933         BUG_ON(!cs->nr_segs);
 934         cs->currbuf = buf;
 935         cs->len = buf->len;
 936         cs->pipebufs++;
 937         cs->nr_segs--;
 938
 939         if (cs->len != PAGE_SIZE)
 940                 goto out_fallback;
 941
 942         if (pipe_buf_steal(cs->pipe, buf) != 0)
 943                 goto out_fallback;
 944
 945         newpage = buf->page;
 946
 947         if (!PageUptodate(newpage))
 948                 SetPageUptodate(newpage);
 949
 950         ClearPageMappedToDisk(newpage);
 951
 952         if (fuse_check_page(newpage) != 0)
 953                 goto out_fallback_unlock;
 954
 955         /*
 956          * This is a new and locked page, it shouldn't be mapped or
 957          * have any special flags on it
 958          */
 959         if (WARN_ON(page_mapped(oldpage)))
 960                 goto out_fallback_unlock;
 961         if (WARN_ON(page_has_private(oldpage)))
 962                 goto out_fallback_unlock;
 963         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 964                 goto out_fallback_unlock;
 965         if (WARN_ON(PageMlocked(oldpage)))
 966                 goto out_fallback_unlock;
 967
 968         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 969         if (err) {
 970                 unlock_page(newpage);
 971                 return err;
 972         }
 973
 974         get_page(newpage);
 975
 976         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 977                 lru_cache_add_file(newpage);
 978
 979         err = 0;
 980         spin_lock(&cs->req->waitq.lock);
 981         if (test_bit(FR_ABORTED, &cs->req->flags))
 982                 err = -ENOENT;
 983         else
 984                 *pagep = newpage;
 985         spin_unlock(&cs->req->waitq.lock);
 986
 987         if (err) {
 988                 unlock_page(newpage);
 989                 put_page(newpage);
 990                 return err;
 991         }
 992
 993         unlock_page(oldpage);
 994         put_page(oldpage);
 995         cs->len = 0;
 996
 997         return 0;
 998
 999 out_fallback_unlock:
1000         unlock_page(newpage);
1001 out_fallback:
1002         cs->pg = buf->page;
1003         cs->offset = buf->offset;
1004
1005         err = lock_request(cs->req);
1006         if (err)
1007                 return err;
1008
1009         return 1;
1010 }
1011
1012 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
1013                          unsigned offset, unsigned count)
1014 {
1015         struct pipe_buffer *buf;
1016         int err;
1017
1018         if (cs->nr_segs == cs->pipe->buffers)
1019                 return -EIO;
1020
1021         err = unlock_request(cs->req);
1022         if (err)
1023                 return err;
1024
1025         fuse_copy_finish(cs);
1026
1027         buf = cs->pipebufs;
1028         get_page(page);
1029         buf->page = page;
1030         buf->offset = offset;
1031         buf->len = count;
1032
1033         cs->pipebufs++;
1034         cs->nr_segs++;
1035         cs->len = 0;
1036
1037         return 0;
1038 }
1039
1040 /*
1041  * Copy a page in the request to/from the userspace buffer.  Must be
1042  * done atomically
1043  */
1044 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
1045                           unsigned offset, unsigned count, int zeroing)
1046 {
1047         int err;
1048         struct page *page = *pagep;
1049
1050         if (page && zeroing && count < PAGE_SIZE)
1051                 clear_highpage(page);
1052
1053         while (count) {
1054                 if (cs->write && cs->pipebufs && page) {
1055                         return fuse_ref_page(cs, page, offset, count);
1056                 } else if (!cs->len) {
1057                         if (cs->move_pages && page &&
1058                             offset == 0 && count == PAGE_SIZE) {
1059                                 err = fuse_try_move_page(cs, pagep);
1060                                 if (err <= 0)
1061                                         return err;
1062                         } else {
1063                                 err = fuse_copy_fill(cs);
1064                                 if (err)
1065                                         return err;
1066                         }
1067                 }
1068                 if (page) {
1069                         void *mapaddr = kmap_atomic(page);
1070                         void *buf = mapaddr + offset;
1071                         offset += fuse_copy_do(cs, &buf, &count);
1072                         kunmap_atomic(mapaddr);
1073                 } else
1074                         offset += fuse_copy_do(cs, NULL, &count);
1075         }
1076         if (page && !cs->write)
1077                 flush_dcache_page(page);
1078         return 0;
1079 }
1080
1081 /* Copy pages in the request to/from userspace buffer */
1082 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1083                            int zeroing)
1084 {
1085         unsigned i;
1086         struct fuse_req *req = cs->req;
1087
1088         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1089                 int err;
1090                 unsigned offset = req->page_descs[i].offset;
1091                 unsigned count = min(nbytes, req->page_descs[i].length);
1092
1093                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1094                                      zeroing);
1095                 if (err)
1096                         return err;
1097
1098                 nbytes -= count;
1099         }
1100         return 0;
1101 }
1102
1103 /* Copy a single argument in the request to/from userspace buffer */
1104 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1105 {
1106         while (size) {
1107                 if (!cs->len) {
1108                         int err = fuse_copy_fill(cs);
1109                         if (err)
1110                                 return err;
1111                 }
1112                 fuse_copy_do(cs, &val, &size);
1113         }
1114         return 0;
1115 }
1116
1117 /* Copy request arguments to/from userspace buffer */
1118 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1119                           unsigned argpages, struct fuse_arg *args,
1120                           int zeroing)
1121 {
1122         int err = 0;
1123         unsigned i;
1124
1125         for (i = 0; !err && i < numargs; i++)  {
1126                 struct fuse_arg *arg = &args[i];
1127                 if (i == numargs - 1 && argpages)
1128                         err = fuse_copy_pages(cs, arg->size, zeroing);
1129                 else
1130                         err = fuse_copy_one(cs, arg->value, arg->size);
1131         }
1132         return err;
1133 }
1134
1135 static int forget_pending(struct fuse_iqueue *fiq)
1136 {
1137         return fiq->forget_list_head.next != NULL;
1138 }
1139
1140 static int request_pending(struct fuse_iqueue *fiq)
1141 {
1142         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1143                 forget_pending(fiq);
1144 }
1145
1146 /*
1147  * Transfer an interrupt request to userspace
1148  *
1149  * Unlike other requests this is assembled on demand, without a need
1150  * to allocate a separate fuse_req structure.
1151  *
1152  * Called with fiq->waitq.lock held, releases it
1153  */
1154 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1155                                struct fuse_copy_state *cs,
1156                                size_t nbytes, struct fuse_req *req)
1157 __releases(fiq->waitq.lock)
1158 {
1159         struct fuse_in_header ih;
1160         struct fuse_interrupt_in arg;
1161         unsigned reqsize = sizeof(ih) + sizeof(arg);
1162         int err;
1163
1164         list_del_init(&req->intr_entry);
1165         memset(&ih, 0, sizeof(ih));
1166         memset(&arg, 0, sizeof(arg));
1167         ih.len = reqsize;
1168         ih.opcode = FUSE_INTERRUPT;
1169         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1170         arg.unique = req->in.h.unique;
1171
1172         spin_unlock(&fiq->waitq.lock);
1173         if (nbytes < reqsize)
1174                 return -EINVAL;
1175
1176         err = fuse_copy_one(cs, &ih, sizeof(ih));
1177         if (!err)
1178                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1179         fuse_copy_finish(cs);
1180
1181         return err ? err : reqsize;
1182 }
1183
1184 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1185                                                unsigned max,
1186                                                unsigned *countp)
1187 {
1188         struct fuse_forget_link *head = fiq->forget_list_head.next;
1189         struct fuse_forget_link **newhead = &head;
1190         unsigned count;
1191
1192         for (count = 0; *newhead != NULL && count < max; count++)
1193                 newhead = &(*newhead)->next;
1194
1195         fiq->forget_list_head.next = *newhead;
1196         *newhead = NULL;
1197         if (fiq->forget_list_head.next == NULL)
1198                 fiq->forget_list_tail = &fiq->forget_list_head;
1199
1200         if (countp != NULL)
1201                 *countp = count;
1202
1203         return head;
1204 }
1205
1206 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1207                                    struct fuse_copy_state *cs,
1208                                    size_t nbytes)
1209 __releases(fiq->waitq.lock)
1210 {
1211         int err;
1212         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1213         struct fuse_forget_in arg = {
1214                 .nlookup = forget->forget_one.nlookup,
1215         };
1216         struct fuse_in_header ih = {
1217                 .opcode = FUSE_FORGET,
1218                 .nodeid = forget->forget_one.nodeid,
1219                 .unique = fuse_get_unique(fiq),
1220                 .len = sizeof(ih) + sizeof(arg),
1221         };
1222
1223         spin_unlock(&fiq->waitq.lock);
1224         kfree(forget);
1225         if (nbytes < ih.len)
1226                 return -EINVAL;
1227
1228         err = fuse_copy_one(cs, &ih, sizeof(ih));
1229         if (!err)
1230                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1231         fuse_copy_finish(cs);
1232
1233         if (err)
1234                 return err;
1235
1236         return ih.len;
1237 }
1238
1239 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1240                                    struct fuse_copy_state *cs, size_t nbytes)
1241 __releases(fiq->waitq.lock)
1242 {
1243         int err;
1244         unsigned max_forgets;
1245         unsigned count;
1246         struct fuse_forget_link *head;
1247         struct fuse_batch_forget_in arg = { .count = 0 };
1248         struct fuse_in_header ih = {
1249                 .opcode = FUSE_BATCH_FORGET,
1250                 .unique = fuse_get_unique(fiq),
1251                 .len = sizeof(ih) + sizeof(arg),
1252         };
1253
1254         if (nbytes < ih.len) {
1255                 spin_unlock(&fiq->waitq.lock);
1256                 return -EINVAL;
1257         }
1258
1259         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1260         head = dequeue_forget(fiq, max_forgets, &count);
1261         spin_unlock(&fiq->waitq.lock);
1262
1263         arg.count = count;
1264         ih.len += count * sizeof(struct fuse_forget_one);
1265         err = fuse_copy_one(cs, &ih, sizeof(ih));
1266         if (!err)
1267                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1268
1269         while (head) {
1270                 struct fuse_forget_link *forget = head;
1271
1272                 if (!err) {
1273                         err = fuse_copy_one(cs, &forget->forget_one,
1274                                             sizeof(forget->forget_one));
1275                 }
1276                 head = forget->next;
1277                 kfree(forget);
1278         }
1279
1280         fuse_copy_finish(cs);
1281
1282         if (err)
1283                 return err;
1284
1285         return ih.len;
1286 }
1287
1288 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1289                             struct fuse_copy_state *cs,
1290                             size_t nbytes)
1291 __releases(fiq->waitq.lock)
1292 {
1293         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1294                 return fuse_read_single_forget(fiq, cs, nbytes);
1295         else
1296                 return fuse_read_batch_forget(fiq, cs, nbytes);
1297 }
1298
1299 /*
1300  * Read a single request into the userspace filesystem's buffer.  This
1301  * function waits until a request is available, then removes it from
1302  * the pending list and copies request data to userspace buffer.  If
1303  * no reply is needed (FORGET) or request has been aborted or there
1304  * was an error during the copying then it's finished by calling
1305  * request_end().  Otherwise add it to the processing list, and set
1306  * the 'sent' flag.
1307  */
1308 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1309                                 struct fuse_copy_state *cs, size_t nbytes)
1310 {
1311         ssize_t err;
1312         struct fuse_conn *fc = fud->fc;
1313         struct fuse_iqueue *fiq = &fc->iq;
1314         struct fuse_pqueue *fpq = &fud->pq;
1315         struct fuse_req *req;
1316         struct fuse_in *in;
1317         unsigned reqsize;
1318         unsigned int hash;
1319
1320         /*
1321          * Require sane minimum read buffer - that has capacity for fixed part
1322          * of any request header + negotated max_write room for data. If the
1323          * requirement is not satisfied return EINVAL to the filesystem server
1324          * to indicate that it is not following FUSE server/client contract.
1325          * Don't dequeue / abort any request.
1326          */
1327         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER, 4096 + fc->max_write))
1328                 return -EINVAL;
1329
1330  restart:
1331         spin_lock(&fiq->waitq.lock);
1332         err = -EAGAIN;
1333         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1334             !request_pending(fiq))
1335                 goto err_unlock;
1336
1337         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1338                                 !fiq->connected || request_pending(fiq));
1339         if (err)
1340                 goto err_unlock;
1341
1342         if (!fiq->connected) {
1343                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1344                 goto err_unlock;
1345         }
1346
1347         if (!list_empty(&fiq->interrupts)) {
1348                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1349                                  intr_entry);
1350                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1351         }
1352
1353         if (forget_pending(fiq)) {
1354                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1355                         return fuse_read_forget(fc, fiq, cs, nbytes);
1356
1357                 if (fiq->forget_batch <= -8)
1358                         fiq->forget_batch = 16;
1359         }
1360
1361         req = list_entry(fiq->pending.next, struct fuse_req, list);
1362         clear_bit(FR_PENDING, &req->flags);
1363         list_del_init(&req->list);
1364         spin_unlock(&fiq->waitq.lock);
1365
1366         in = &req->in;
1367         reqsize = in->h.len;
1368
1369         /* If request is too large, reply with an error and restart the read */
1370         if (nbytes < reqsize) {
1371                 req->out.h.error = -EIO;
1372                 /* SETXATTR is special, since it may contain too large data */
1373                 if (in->h.opcode == FUSE_SETXATTR)
1374                         req->out.h.error = -E2BIG;
1375                 request_end(fc, req);
1376                 goto restart;
1377         }
1378         spin_lock(&fpq->lock);
1379         list_add(&req->list, &fpq->io);
1380         spin_unlock(&fpq->lock);
1381         cs->req = req;
1382         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1383         if (!err)
1384                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1385                                      (struct fuse_arg *) in->args, 0);
1386         fuse_copy_finish(cs);
1387         spin_lock(&fpq->lock);
1388         clear_bit(FR_LOCKED, &req->flags);
1389         if (!fpq->connected) {
1390                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1391                 goto out_end;
1392         }
1393         if (err) {
1394                 req->out.h.error = -EIO;
1395                 goto out_end;
1396         }
1397         if (!test_bit(FR_ISREPLY, &req->flags)) {
1398                 err = reqsize;
1399                 goto out_end;
1400         }
1401         hash = fuse_req_hash(req->in.h.unique);
1402         list_move_tail(&req->list, &fpq->processing[hash]);
1403         __fuse_get_request(req);
1404         set_bit(FR_SENT, &req->flags);
1405         spin_unlock(&fpq->lock);
1406         /* matches barrier in request_wait_answer() */
1407         smp_mb__after_atomic();
1408         if (test_bit(FR_INTERRUPTED, &req->flags))
1409                 queue_interrupt(fiq, req);
1410         fuse_put_request(fc, req);
1411
1412         return reqsize;
1413
1414 out_end:
1415         if (!test_bit(FR_PRIVATE, &req->flags))
1416                 list_del_init(&req->list);
1417         spin_unlock(&fpq->lock);
1418         request_end(fc, req);
1419         return err;
1420
1421  err_unlock:
1422         spin_unlock(&fiq->waitq.lock);
1423         return err;
1424 }
1425
1426 static int fuse_dev_open(struct inode *inode, struct file *file)
1427 {
1428         /*
1429          * The fuse device's file's private_data is used to hold
1430          * the fuse_conn(ection) when it is mounted, and is used to
1431          * keep track of whether the file has been mounted already.
1432          */
1433         file->private_data = NULL;
1434         return 0;
1435 }
1436
1437 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1438 {
1439         struct fuse_copy_state cs;
1440         struct file *file = iocb->ki_filp;
1441         struct fuse_dev *fud = fuse_get_dev(file);
1442
1443         if (!fud)
1444                 return -EPERM;
1445
1446         if (!iter_is_iovec(to))
1447                 return -EINVAL;
1448
1449         fuse_copy_init(&cs, 1, to);
1450
1451         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1452 }
1453
1454 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1455                                     struct pipe_inode_info *pipe,
1456                                     size_t len, unsigned int flags)
1457 {
1458         int total, ret;
1459         int page_nr = 0;
1460         struct pipe_buffer *bufs;
1461         struct fuse_copy_state cs;
1462         struct fuse_dev *fud = fuse_get_dev(in);
1463
1464         if (!fud)
1465                 return -EPERM;
1466
1467         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1468                               GFP_KERNEL);
1469         if (!bufs)
1470                 return -ENOMEM;
1471
1472         fuse_copy_init(&cs, 1, NULL);
1473         cs.pipebufs = bufs;
1474         cs.pipe = pipe;
1475         ret = fuse_dev_do_read(fud, in, &cs, len);
1476         if (ret < 0)
1477                 goto out;
1478
1479         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1480                 ret = -EIO;
1481                 goto out;
1482         }
1483
1484         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1485                 /*
1486                  * Need to be careful about this.  Having buf->ops in module
1487                  * code can Oops if the buffer persists after module unload.
1488                  */
1489                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1490                 bufs[page_nr].flags = 0;
1491                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1492                 if (unlikely(ret < 0))
1493                         break;
1494         }
1495         if (total)
1496                 ret = total;
1497 out:
1498         for (; page_nr < cs.nr_segs; page_nr++)
1499                 put_page(bufs[page_nr].page);
1500
1501         kvfree(bufs);
1502         return ret;
1503 }
1504
1505 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1506                             struct fuse_copy_state *cs)
1507 {
1508         struct fuse_notify_poll_wakeup_out outarg;
1509         int err = -EINVAL;
1510
1511         if (size != sizeof(outarg))
1512                 goto err;
1513
1514         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1515         if (err)
1516                 goto err;
1517
1518         fuse_copy_finish(cs);
1519         return fuse_notify_poll_wakeup(fc, &outarg);
1520
1521 err:
1522         fuse_copy_finish(cs);
1523         return err;
1524 }
1525
1526 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1527                                    struct fuse_copy_state *cs)
1528 {
1529         struct fuse_notify_inval_inode_out outarg;
1530         int err = -EINVAL;
1531
1532         if (size != sizeof(outarg))
1533                 goto err;
1534
1535         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1536         if (err)
1537                 goto err;
1538         fuse_copy_finish(cs);
1539
1540         down_read(&fc->killsb);
1541         err = -ENOENT;
1542         if (fc->sb) {
1543                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1544                                                outarg.off, outarg.len);
1545         }
1546         up_read(&fc->killsb);
1547         return err;
1548
1549 err:
1550         fuse_copy_finish(cs);
1551         return err;
1552 }
1553
1554 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1555                                    struct fuse_copy_state *cs)
1556 {
1557         struct fuse_notify_inval_entry_out outarg;
1558         int err = -ENOMEM;
1559         char *buf;
1560         struct qstr name;
1561
1562         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1563         if (!buf)
1564                 goto err;
1565
1566         err = -EINVAL;
1567         if (size < sizeof(outarg))
1568                 goto err;
1569
1570         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1571         if (err)
1572                 goto err;
1573
1574         err = -ENAMETOOLONG;
1575         if (outarg.namelen > FUSE_NAME_MAX)
1576                 goto err;
1577
1578         err = -EINVAL;
1579         if (size != sizeof(outarg) + outarg.namelen + 1)
1580                 goto err;
1581
1582         name.name = buf;
1583         name.len = outarg.namelen;
1584         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1585         if (err)
1586                 goto err;
1587         fuse_copy_finish(cs);
1588         buf[outarg.namelen] = 0;
1589
1590         down_read(&fc->killsb);
1591         err = -ENOENT;
1592         if (fc->sb)
1593                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1594         up_read(&fc->killsb);
1595         kfree(buf);
1596         return err;
1597
1598 err:
1599         kfree(buf);
1600         fuse_copy_finish(cs);
1601         return err;
1602 }
1603
1604 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1605                               struct fuse_copy_state *cs)
1606 {
1607         struct fuse_notify_delete_out outarg;
1608         int err = -ENOMEM;
1609         char *buf;
1610         struct qstr name;
1611
1612         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1613         if (!buf)
1614                 goto err;
1615
1616         err = -EINVAL;
1617         if (size < sizeof(outarg))
1618                 goto err;
1619
1620         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1621         if (err)
1622                 goto err;
1623
1624         err = -ENAMETOOLONG;
1625         if (outarg.namelen > FUSE_NAME_MAX)
1626                 goto err;
1627
1628         err = -EINVAL;
1629         if (size != sizeof(outarg) + outarg.namelen + 1)
1630                 goto err;
1631
1632         name.name = buf;
1633         name.len = outarg.namelen;
1634         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1635         if (err)
1636                 goto err;
1637         fuse_copy_finish(cs);
1638         buf[outarg.namelen] = 0;
1639
1640         down_read(&fc->killsb);
1641         err = -ENOENT;
1642         if (fc->sb)
1643                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1644                                                outarg.child, &name);
1645         up_read(&fc->killsb);
1646         kfree(buf);
1647         return err;
1648
1649 err:
1650         kfree(buf);
1651         fuse_copy_finish(cs);
1652         return err;
1653 }
1654
1655 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1656                              struct fuse_copy_state *cs)
1657 {
1658         struct fuse_notify_store_out outarg;
1659         struct inode *inode;
1660         struct address_space *mapping;
1661         u64 nodeid;
1662         int err;
1663         pgoff_t index;
1664         unsigned int offset;
1665         unsigned int num;
1666         loff_t file_size;
1667         loff_t end;
1668
1669         err = -EINVAL;
1670         if (size < sizeof(outarg))
1671                 goto out_finish;
1672
1673         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1674         if (err)
1675                 goto out_finish;
1676
1677         err = -EINVAL;
1678         if (size - sizeof(outarg) != outarg.size)
1679                 goto out_finish;
1680
1681         nodeid = outarg.nodeid;
1682
1683         down_read(&fc->killsb);
1684
1685         err = -ENOENT;
1686         if (!fc->sb)
1687                 goto out_up_killsb;
1688
1689         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1690         if (!inode)
1691                 goto out_up_killsb;
1692
1693         mapping = inode->i_mapping;
1694         index = outarg.offset >> PAGE_SHIFT;
1695         offset = outarg.offset & ~PAGE_MASK;
1696         file_size = i_size_read(inode);
1697         end = outarg.offset + outarg.size;
1698         if (end > file_size) {
1699                 file_size = end;
1700                 fuse_write_update_size(inode, file_size);
1701         }
1702
1703         num = outarg.size;
1704         while (num) {
1705                 struct page *page;
1706                 unsigned int this_num;
1707
1708                 err = -ENOMEM;
1709                 page = find_or_create_page(mapping, index,
1710                                            mapping_gfp_mask(mapping));
1711                 if (!page)
1712                         goto out_iput;
1713
1714                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1715                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1716                 if (!err && offset == 0 &&
1717                     (this_num == PAGE_SIZE || file_size == end))
1718                         SetPageUptodate(page);
1719                 unlock_page(page);
1720                 put_page(page);
1721
1722                 if (err)
1723                         goto out_iput;
1724
1725                 num -= this_num;
1726                 offset = 0;
1727                 index++;
1728         }
1729
1730         err = 0;
1731
1732 out_iput:
1733         iput(inode);
1734 out_up_killsb:
1735         up_read(&fc->killsb);
1736 out_finish:
1737         fuse_copy_finish(cs);
1738         return err;
1739 }
1740
1741 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1742 {
1743         release_pages(req->pages, req->num_pages);
1744 }
1745
1746 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1747                          struct fuse_notify_retrieve_out *outarg)
1748 {
1749         int err;
1750         struct address_space *mapping = inode->i_mapping;
1751         struct fuse_req *req;
1752         pgoff_t index;
1753         loff_t file_size;
1754         unsigned int num;
1755         unsigned int offset;
1756         size_t total_len = 0;
1757         unsigned int num_pages;
1758
1759         offset = outarg->offset & ~PAGE_MASK;
1760         file_size = i_size_read(inode);
1761
1762         num = min(outarg->size, fc->max_write);
1763         if (outarg->offset > file_size)
1764                 num = 0;
1765         else if (outarg->offset + num > file_size)
1766                 num = file_size - outarg->offset;
1767
1768         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1769         num_pages = min(num_pages, fc->max_pages);
1770
1771         req = fuse_get_req(fc, num_pages);
1772         if (IS_ERR(req))
1773                 return PTR_ERR(req);
1774
1775         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1776         req->in.h.nodeid = outarg->nodeid;
1777         req->in.numargs = 2;
1778         req->in.argpages = 1;
1779         req->end = fuse_retrieve_end;
1780
1781         index = outarg->offset >> PAGE_SHIFT;
1782
1783         while (num && req->num_pages < num_pages) {
1784                 struct page *page;
1785                 unsigned int this_num;
1786
1787                 page = find_get_page(mapping, index);
1788                 if (!page)
1789                         break;
1790
1791                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1792                 req->pages[req->num_pages] = page;
1793                 req->page_descs[req->num_pages].offset = offset;
1794                 req->page_descs[req->num_pages].length = this_num;
1795                 req->num_pages++;
1796
1797                 offset = 0;
1798                 num -= this_num;
1799                 total_len += this_num;
1800                 index++;
1801         }
1802         req->misc.retrieve_in.offset = outarg->offset;
1803         req->misc.retrieve_in.size = total_len;
1804         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1805         req->in.args[0].value = &req->misc.retrieve_in;
1806         req->in.args[1].size = total_len;
1807
1808         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1809         if (err) {
1810                 fuse_retrieve_end(fc, req);
1811                 fuse_put_request(fc, req);
1812         }
1813
1814         return err;
1815 }
1816
1817 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1818                                 struct fuse_copy_state *cs)
1819 {
1820         struct fuse_notify_retrieve_out outarg;
1821         struct inode *inode;
1822         int err;
1823
1824         err = -EINVAL;
1825         if (size != sizeof(outarg))
1826                 goto copy_finish;
1827
1828         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1829         if (err)
1830                 goto copy_finish;
1831
1832         fuse_copy_finish(cs);
1833
1834         down_read(&fc->killsb);
1835         err = -ENOENT;
1836         if (fc->sb) {
1837                 u64 nodeid = outarg.nodeid;
1838
1839                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1840                 if (inode) {
1841                         err = fuse_retrieve(fc, inode, &outarg);
1842                         iput(inode);
1843                 }
1844         }
1845         up_read(&fc->killsb);
1846
1847         return err;
1848
1849 copy_finish:
1850         fuse_copy_finish(cs);
1851         return err;
1852 }
1853
1854 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1855                        unsigned int size, struct fuse_copy_state *cs)
1856 {
1857         /* Don't try to move pages (yet) */
1858         cs->move_pages = 0;
1859
1860         switch (code) {
1861         case FUSE_NOTIFY_POLL:
1862                 return fuse_notify_poll(fc, size, cs);
1863
1864         case FUSE_NOTIFY_INVAL_INODE:
1865                 return fuse_notify_inval_inode(fc, size, cs);
1866
1867         case FUSE_NOTIFY_INVAL_ENTRY:
1868                 return fuse_notify_inval_entry(fc, size, cs);
1869
1870         case FUSE_NOTIFY_STORE:
1871                 return fuse_notify_store(fc, size, cs);
1872
1873         case FUSE_NOTIFY_RETRIEVE:
1874                 return fuse_notify_retrieve(fc, size, cs);
1875
1876         case FUSE_NOTIFY_DELETE:
1877                 return fuse_notify_delete(fc, size, cs);
1878
1879         default:
1880                 fuse_copy_finish(cs);
1881                 return -EINVAL;
1882         }
1883 }
1884
1885 /* Look up request on processing list by unique ID */
1886 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1887 {
1888         unsigned int hash = fuse_req_hash(unique);
1889         struct fuse_req *req;
1890
1891         list_for_each_entry(req, &fpq->processing[hash], list) {
1892                 if (req->in.h.unique == unique)
1893                         return req;
1894         }
1895         return NULL;
1896 }
1897
1898 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1899                          unsigned nbytes)
1900 {
1901         unsigned reqsize = sizeof(struct fuse_out_header);
1902
1903         if (out->h.error)
1904                 return nbytes != reqsize ? -EINVAL : 0;
1905
1906         reqsize += len_args(out->numargs, out->args);
1907
1908         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1909                 return -EINVAL;
1910         else if (reqsize > nbytes) {
1911                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1912                 unsigned diffsize = reqsize - nbytes;
1913                 if (diffsize > lastarg->size)
1914                         return -EINVAL;
1915                 lastarg->size -= diffsize;
1916         }
1917         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1918                               out->page_zeroing);
1919 }
1920
1921 /*
1922  * Write a single reply to a request.  First the header is copied from
1923  * the write buffer.  The request is then searched on the processing
1924  * list by the unique ID found in the header.  If found, then remove
1925  * it from the list and copy the rest of the buffer to the request.
1926  * The request is finished by calling request_end()
1927  */
1928 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1929                                  struct fuse_copy_state *cs, size_t nbytes)
1930 {
1931         int err;
1932         struct fuse_conn *fc = fud->fc;
1933         struct fuse_pqueue *fpq = &fud->pq;
1934         struct fuse_req *req;
1935         struct fuse_out_header oh;
1936
1937         err = -EINVAL;
1938         if (nbytes < sizeof(struct fuse_out_header))
1939                 goto out;
1940
1941         err = fuse_copy_one(cs, &oh, sizeof(oh));
1942         if (err)
1943                 goto copy_finish;
1944
1945         err = -EINVAL;
1946         if (oh.len != nbytes)
1947                 goto copy_finish;
1948
1949         /*
1950          * Zero oh.unique indicates unsolicited notification message
1951          * and error contains notification code.
1952          */
1953         if (!oh.unique) {
1954                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1955                 goto out;
1956         }
1957
1958         err = -EINVAL;
1959         if (oh.error <= -1000 || oh.error > 0)
1960                 goto copy_finish;
1961
1962         spin_lock(&fpq->lock);
1963         req = NULL;
1964         if (fpq->connected)
1965                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1966
1967         err = -ENOENT;
1968         if (!req) {
1969                 spin_unlock(&fpq->lock);
1970                 goto copy_finish;
1971         }
1972
1973         /* Is it an interrupt reply ID? */
1974         if (oh.unique & FUSE_INT_REQ_BIT) {
1975                 __fuse_get_request(req);
1976                 spin_unlock(&fpq->lock);
1977
1978                 err = 0;
1979                 if (nbytes != sizeof(struct fuse_out_header))
1980                         err = -EINVAL;
1981                 else if (oh.error == -ENOSYS)
1982                         fc->no_interrupt = 1;
1983                 else if (oh.error == -EAGAIN)
1984                         err = queue_interrupt(&fc->iq, req);
1985
1986                 fuse_put_request(fc, req);
1987
1988                 goto copy_finish;
1989         }
1990
1991         clear_bit(FR_SENT, &req->flags);
1992         list_move(&req->list, &fpq->io);
1993         req->out.h = oh;
1994         set_bit(FR_LOCKED, &req->flags);
1995         spin_unlock(&fpq->lock);
1996         cs->req = req;
1997         if (!req->out.page_replace)
1998                 cs->move_pages = 0;
1999
2000         err = copy_out_args(cs, &req->out, nbytes);
2001         fuse_copy_finish(cs);
2002
2003         spin_lock(&fpq->lock);
2004         clear_bit(FR_LOCKED, &req->flags);
2005         if (!fpq->connected)
2006                 err = -ENOENT;
2007         else if (err)
2008                 req->out.h.error = -EIO;
2009         if (!test_bit(FR_PRIVATE, &req->flags))
2010                 list_del_init(&req->list);
2011         spin_unlock(&fpq->lock);
2012
2013         request_end(fc, req);
2014 out:
2015         return err ? err : nbytes;
2016
2017 copy_finish:
2018         fuse_copy_finish(cs);
2019         goto out;
2020 }
2021
2022 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
2023 {
2024         struct fuse_copy_state cs;
2025         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
2026
2027         if (!fud)
2028                 return -EPERM;
2029
2030         if (!iter_is_iovec(from))
2031                 return -EINVAL;
2032
2033         fuse_copy_init(&cs, 0, from);
2034
2035         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2036 }
2037
2038 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2039                                      struct file *out, loff_t *ppos,
2040                                      size_t len, unsigned int flags)
2041 {
2042         unsigned nbuf;
2043         unsigned idx;
2044         struct pipe_buffer *bufs;
2045         struct fuse_copy_state cs;
2046         struct fuse_dev *fud;
2047         size_t rem;
2048         ssize_t ret;
2049
2050         fud = fuse_get_dev(out);
2051         if (!fud)
2052                 return -EPERM;
2053
2054         pipe_lock(pipe);
2055
2056         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
2057                               GFP_KERNEL);
2058         if (!bufs) {
2059                 pipe_unlock(pipe);
2060                 return -ENOMEM;
2061         }
2062
2063         nbuf = 0;
2064         rem = 0;
2065         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2066                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2067
2068         ret = -EINVAL;
2069         if (rem < len)
2070                 goto out_free;
2071
2072         rem = len;
2073         while (rem) {
2074                 struct pipe_buffer *ibuf;
2075                 struct pipe_buffer *obuf;
2076
2077                 BUG_ON(nbuf >= pipe->buffers);
2078                 BUG_ON(!pipe->nrbufs);
2079                 ibuf = &pipe->bufs[pipe->curbuf];
2080                 obuf = &bufs[nbuf];
2081
2082                 if (rem >= ibuf->len) {
2083                         *obuf = *ibuf;
2084                         ibuf->ops = NULL;
2085                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2086                         pipe->nrbufs--;
2087                 } else {
2088                         if (!pipe_buf_get(pipe, ibuf))
2089                                 goto out_free;
2090
2091                         *obuf = *ibuf;
2092                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2093                         obuf->len = rem;
2094                         ibuf->offset += obuf->len;
2095                         ibuf->len -= obuf->len;
2096                 }
2097                 nbuf++;
2098                 rem -= obuf->len;
2099         }
2100         pipe_unlock(pipe);
2101
2102         fuse_copy_init(&cs, 0, NULL);
2103         cs.pipebufs = bufs;
2104         cs.nr_segs = nbuf;
2105         cs.pipe = pipe;
2106
2107         if (flags & SPLICE_F_MOVE)
2108                 cs.move_pages = 1;
2109
2110         ret = fuse_dev_do_write(fud, &cs, len);
2111
2112         pipe_lock(pipe);
2113 out_free:
2114         for (idx = 0; idx < nbuf; idx++)
2115                 pipe_buf_release(pipe, &bufs[idx]);
2116         pipe_unlock(pipe);
2117
2118         kvfree(bufs);
2119         return ret;
2120 }
2121
2122 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2123 {
2124         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2125         struct fuse_iqueue *fiq;
2126         struct fuse_dev *fud = fuse_get_dev(file);
2127
2128         if (!fud)
2129                 return EPOLLERR;
2130
2131         fiq = &fud->fc->iq;
2132         poll_wait(file, &fiq->waitq, wait);
2133
2134         spin_lock(&fiq->waitq.lock);
2135         if (!fiq->connected)
2136                 mask = EPOLLERR;
2137         else if (request_pending(fiq))
2138                 mask |= EPOLLIN | EPOLLRDNORM;
2139         spin_unlock(&fiq->waitq.lock);
2140
2141         return mask;
2142 }
2143
2144 /* Abort all requests on the given list (pending or processing) */
2145 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2146 {
2147         while (!list_empty(head)) {
2148                 struct fuse_req *req;
2149                 req = list_entry(head->next, struct fuse_req, list);
2150                 req->out.h.error = -ECONNABORTED;
2151                 clear_bit(FR_SENT, &req->flags);
2152                 list_del_init(&req->list);
2153                 request_end(fc, req);
2154         }
2155 }
2156
2157 static void end_polls(struct fuse_conn *fc)
2158 {
2159         struct rb_node *p;
2160
2161         p = rb_first(&fc->polled_files);
2162
2163         while (p) {
2164                 struct fuse_file *ff;
2165                 ff = rb_entry(p, struct fuse_file, polled_node);
2166                 wake_up_interruptible_all(&ff->poll_wait);
2167
2168                 p = rb_next(p);
2169         }
2170 }
2171
2172 /*
2173  * Abort all requests.
2174  *
2175  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2176  * filesystem.
2177  *
2178  * The same effect is usually achievable through killing the filesystem daemon
2179  * and all users of the filesystem.  The exception is the combination of an
2180  * asynchronous request and the tricky deadlock (see
2181  * Documentation/filesystems/fuse.txt).
2182  *
2183  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2184  * requests, they should be finished off immediately.  Locked requests will be
2185  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2186  * requests.  It is possible that some request will finish before we can.  This
2187  * is OK, the request will in that case be removed from the list before we touch
2188  * it.
2189  */
2190 void fuse_abort_conn(struct fuse_conn *fc)
2191 {
2192         struct fuse_iqueue *fiq = &fc->iq;
2193
2194         spin_lock(&fc->lock);
2195         if (fc->connected) {
2196                 struct fuse_dev *fud;
2197                 struct fuse_req *req, *next;
2198                 LIST_HEAD(to_end);
2199                 unsigned int i;
2200
2201                 /* Background queuing checks fc->connected under bg_lock */
2202                 spin_lock(&fc->bg_lock);
2203                 fc->connected = 0;
2204                 spin_unlock(&fc->bg_lock);
2205
2206                 fuse_set_initialized(fc);
2207                 list_for_each_entry(fud, &fc->devices, entry) {
2208                         struct fuse_pqueue *fpq = &fud->pq;
2209
2210                         spin_lock(&fpq->lock);
2211                         fpq->connected = 0;
2212                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2213                                 req->out.h.error = -ECONNABORTED;
2214                                 spin_lock(&req->waitq.lock);
2215                                 set_bit(FR_ABORTED, &req->flags);
2216                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2217                                         set_bit(FR_PRIVATE, &req->flags);
2218                                         __fuse_get_request(req);
2219                                         list_move(&req->list, &to_end);
2220                                 }
2221                                 spin_unlock(&req->waitq.lock);
2222                         }
2223                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2224                                 list_splice_tail_init(&fpq->processing[i],
2225                                                       &to_end);
2226                         spin_unlock(&fpq->lock);
2227                 }
2228                 spin_lock(&fc->bg_lock);
2229                 fc->blocked = 0;
2230                 fc->max_background = UINT_MAX;
2231                 flush_bg_queue(fc);
2232                 spin_unlock(&fc->bg_lock);
2233
2234                 spin_lock(&fiq->waitq.lock);
2235                 fiq->connected = 0;
2236                 list_for_each_entry(req, &fiq->pending, list)
2237                         clear_bit(FR_PENDING, &req->flags);
2238                 list_splice_tail_init(&fiq->pending, &to_end);
2239                 while (forget_pending(fiq))
2240                         kfree(dequeue_forget(fiq, 1, NULL));
2241                 wake_up_all_locked(&fiq->waitq);
2242                 spin_unlock(&fiq->waitq.lock);
2243                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2244                 end_polls(fc);
2245                 wake_up_all(&fc->blocked_waitq);
2246                 spin_unlock(&fc->lock);
2247
2248                 end_requests(fc, &to_end);
2249         } else {
2250                 spin_unlock(&fc->lock);
2251         }
2252 }
2253 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2254
2255 void fuse_wait_aborted(struct fuse_conn *fc)
2256 {
2257         /* matches implicit memory barrier in fuse_drop_waiting() */
2258         smp_mb();
2259         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2260 }
2261
2262 int fuse_dev_release(struct inode *inode, struct file *file)
2263 {
2264         struct fuse_dev *fud = fuse_get_dev(file);
2265
2266         if (fud) {
2267                 struct fuse_conn *fc = fud->fc;
2268                 struct fuse_pqueue *fpq = &fud->pq;
2269                 LIST_HEAD(to_end);
2270                 unsigned int i;
2271
2272                 spin_lock(&fpq->lock);
2273                 WARN_ON(!list_empty(&fpq->io));
2274                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2275                         list_splice_init(&fpq->processing[i], &to_end);
2276                 spin_unlock(&fpq->lock);
2277
2278                 end_requests(fc, &to_end);
2279
2280                 /* Are we the last open device? */
2281                 if (atomic_dec_and_test(&fc->dev_count)) {
2282                         WARN_ON(fc->iq.fasync != NULL);
2283                         fuse_abort_conn(fc);
2284                 }
2285                 fuse_dev_free(fud);
2286         }
2287         return 0;
2288 }
2289 EXPORT_SYMBOL_GPL(fuse_dev_release);
2290
2291 static int fuse_dev_fasync(int fd, struct file *file, int on)
2292 {
2293         struct fuse_dev *fud = fuse_get_dev(file);
2294
2295         if (!fud)
2296                 return -EPERM;
2297
2298         /* No locking - fasync_helper does its own locking */
2299         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2300 }
2301
2302 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2303 {
2304         struct fuse_dev *fud;
2305
2306         if (new->private_data)
2307                 return -EINVAL;
2308
2309         fud = fuse_dev_alloc(fc);
2310         if (!fud)
2311                 return -ENOMEM;
2312
2313         new->private_data = fud;
2314         atomic_inc(&fc->dev_count);
2315
2316         return 0;
2317 }
2318
2319 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2320                            unsigned long arg)
2321 {
2322         int err = -ENOTTY;
2323
2324         if (cmd == FUSE_DEV_IOC_CLONE) {
2325                 int oldfd;
2326
2327                 err = -EFAULT;
2328                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2329                         struct file *old = fget(oldfd);
2330
2331                         err = -EINVAL;
2332                         if (old) {
2333                                 struct fuse_dev *fud = NULL;
2334
2335                                 /*
2336                                  * Check against file->f_op because CUSE
2337                                  * uses the same ioctl handler.
2338                                  */
2339                                 if (old->f_op == file->f_op &&
2340                                     old->f_cred->user_ns == file->f_cred->user_ns)
2341                                         fud = fuse_get_dev(old);
2342
2343                                 if (fud) {
2344                                         mutex_lock(&fuse_mutex);
2345                                         err = fuse_device_clone(fud->fc, file);
2346                                         mutex_unlock(&fuse_mutex);
2347                                 }
2348                                 fput(old);
2349                         }
2350                 }
2351         }
2352         return err;
2353 }
2354
2355 const struct file_operations fuse_dev_operations = {
2356         .owner          = THIS_MODULE,
2357         .open           = fuse_dev_open,
2358         .llseek         = no_llseek,
2359         .read_iter      = fuse_dev_read,
2360         .splice_read    = fuse_dev_splice_read,
2361         .write_iter     = fuse_dev_write,
2362         .splice_write   = fuse_dev_splice_write,
2363         .poll           = fuse_dev_poll,
2364         .release        = fuse_dev_release,
2365         .fasync         = fuse_dev_fasync,
2366         .unlocked_ioctl = fuse_dev_ioctl,
2367         .compat_ioctl   = fuse_dev_ioctl,
2368 };
2369 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2370
2371 static struct miscdevice fuse_miscdevice = {
2372         .minor = FUSE_MINOR,
2373         .name  = "fuse",
2374         .fops = &fuse_dev_operations,
2375 };
2376
2377 int __init fuse_dev_init(void)
2378 {
2379         int err = -ENOMEM;
2380         fuse_req_cachep = kmem_cache_create("fuse_request",
2381                                             sizeof(struct fuse_req),
2382                                             0, 0, NULL);
2383         if (!fuse_req_cachep)
2384                 goto out;
2385
2386         err = misc_register(&fuse_miscdevice);
2387         if (err)
2388                 goto out_cache_clean;
2389
2390         return 0;
2391
2392  out_cache_clean:
2393         kmem_cache_destroy(fuse_req_cachep);
2394  out:
2395         return err;
2396 }
2397
2398 void fuse_dev_cleanup(void)
2399 {
2400         misc_deregister(&fuse_miscdevice);
2401         kmem_cache_destroy(fuse_req_cachep);
2402 }